cerredz Multi-Agent Debate Reasoning Trace
Goal
Produce a complete, timestamped audit trail of a Multi-Agent Debate reasoning session that captures every agent position, targeted critique, accepted or rejected revision, and convergence or divergence outcome across all three debate rounds. The trace simulates heterogeneous agents with distinct cognitive rolesΓÇöOptimizer, Skeptic, and IntegratorΓÇöwho independently apply a selected Vidbyte reasoning strategy to the user's problem, then engage in structured adversarial collaboration where each agent's claims must survive targeted cross-examination before earning a place in the final answer. The trace serves as both a quality assurance artifactΓÇöallowing a human reviewer to independently verify which claims survived scrutiny and which were revised or abandonedΓÇöand a feedback mechanism for improving the debate architecture itself by identifying patterns in critique effectiveness, convergence rate, and role differentiation quality across sessions.
The trace is designed to be both human-readable and machine-parseable, structured as markdown with clearly delineated sections for each round and agent, so it can feed into automated analysis pipelines, regression suites for strategy output robustness, and continuous improvement loops where debate-pattern analytics surface which agent role configurations and critique strategies most reliably produce convergence on correct answers.
Intent
The purpose of this trace is to give the model a concrete mechanism for implementing the multi-agent debate techniqueΓÇöstructured adversarial collaboration between distinct perspectives that improves factuality and logical consistency. Without a trace specification, the model may produce a superficial debate that looks structured on the surface but lacks genuine adversarial pressure: agents may agree prematurely, critiques may be generic, and disagreements may be papered over rather than resolved or characterized. This trace enforces the discipline of heterogeneous role instantiation, forced critique specificity, transparent revision tracking, and honest convergence assessment. By requiring each agent to produce distinct output with its assigned epistemic posture, requiring critiques to reference specific claims, and requiring the Integrator to map consensus, partial-consensus, and disagreement zones explicitly, the trace transforms debate from a narrative device into a verifiable reasoning instrument.
Instructions
Derive {question_name} from the user's actual question by lowercasing it, replacing non-alphanumeric runs with hyphens, trimming extra hyphens, and using reasoning-trace if no safe name remains. Create the root memory directory when needed, then write or replace memory/{question_name}.md with this trace. Start the file with the question, selected strategy, scale note, source constraints, and a brief statement of what the trace will inspect. Build the scratchpad by repeatedly applying the multi-agent debate move: instantiate three heterogeneous agents, collect Round 1 independent positions, execute Round 2 structured critiques, execute Round 3 revisions with accept/reject annotations, and characterize convergence. Use a medium-depth trace by default, usually around 150 numbered lines, unless the user asks for a different depth. Prefer concise public reasoning artifacts over hidden deliberation, and make every numbered item contribute a position, critique, revision, or convergence assessment. End the file with a synthesis and final answer that follow from the trace, including any important disagreement or uncertainty that remains.
Background Information
Multi-Agent Debate reasoning, formalized by Du et al. (2023) in "Improving Factuality and Reasoning through Multiagent Debate," operationalizes the Society of Mind principle that intelligence emerges from the interaction of specialized agents rather than from a single monolithic reasoner. In the MAD framework, multiple language model instances independently generate answers to a question, then engage in structured debate rounds where each agent sees the others' responses and produces critiques and revisions. Successive rounds drive convergence toward correct answers through a selection pressure: claims that survive adversarial scrutiny persist, while claims that cannot withstand targeted critique are revised or abandoned. Du et al. demonstrated that MAD improves mathematical reasoning accuracy on GSM8K and factual accuracy on truthfulness benchmarks, with debate rounds producing measurable convergence toward correct answers.
The central insight of the MAD framework is that heterogeneous agentsΓÇöthose with genuinely distinct cognitive roles, reasoning styles, or epistemic prioritiesΓÇöoutperform homogeneous agents whose similar biases produce correlated errors that majority voting cannot correct. When all agents share the same blind spots, debate devolves into mutual reinforcement rather than genuine testing. Heterogeneous roles create the cognitive diversity necessary for critique to surface weaknesses that any single perspective would miss. The debate architecture records disagreements transparently rather than forcing consensus, ensuring that unresolved tensions are visible to the user rather than hidden behind a facade of agreement. For this trace, the three canonical heterogeneous roles are the Optimizer (prioritizing solution strength and completeness), the Skeptic (prioritizing flaw detection and alternative explanations), and the Integrator (prioritizing coherence, reconciliation, and honest uncertainty characterization). Each role applies the selected Vidbyte strategy through its own epistemic lens, producing positions that differ not in correctness but in emphasis, risk posture, and weighting of competing considerations.
Algorithm for the Output Document
Restate the user's question, constraints, evidence standard, and strategy domain, then derive the subquestions or working items that the selected strategy needs to inspect. Confirm that all three agent role definitions are loaded and that each has a clearly differentiated epistemic posture.
Define the three heterogeneous agents explicitly in the trace:
- Optimizer: Produces the strongest, most complete application of the selected strategyΓÇöprioritizes solution quality, comprehensiveness, and actionability. Pushes for the best possible answer with maximal supporting reasoning and concrete recommendations.
- Skeptic: Finds every weakness, gap, hidden assumption, alternative explanation, and failure mode in others' positionsΓÇöprioritizes critical scrutiny, demanding evidence for every claim and surfacing counterarguments that others missed. Accepts no claim at face value.
- Integrator: Reconciles competing positions, identifies where agents actually agree versus where they genuinely differ, synthesizes partial agreements into a coherent whole, and characterizes residual disagreements clearlyΓÇöprioritizes coherence, completeness, and honest uncertainty representation.
Execute Round 1 ΓÇö Initial Positions: Each agent independently applies the selected reasoning strategy to the user's problem and produces a complete initial position. Number each position item within the round. Record each agent's full output with an explicit confidence level for its overall position and for key claims within it. Ensure that the three positions are recognizably distinct in emphasis and risk postureΓÇöthe Optimizer should not read like the Skeptic, and vice versa.
Execute Round 2 ΓÇö Structured Critique: Each agent reviews the other two positions and produces targeted critiques. Every critique must reference a specific numbered claim from the target agent's position; generic criticism without a specific target is invalid and must be discarded. The Optimizer critiques for missed opportunities, suboptimal solutions, insufficient ambition, and failure to maximize strategic value. The Skeptic critiques for logical gaps, unsupported claims, hidden assumptions, missing evidence, alternative explanations, and failure modes the target agent did not consider. The Integrator critiques for internal inconsistency, unclear tradeoffs, unresolved tensions between competing claims, and failure to address the full scope of the user's problem. Record each critique with a cross-reference to the claim it targets.
Execute Round 3 ΓÇö Revision and Convergence: Each agent revises its position in light of the critiques received. For every critique, the revising agent must explicitly state whether it accepted the critique (and how it modified its position), rejected the critique (and why the critique was unfounded or the agent's original position was correct), or partially accepted the critique (what changed and what did not). Number each revision item and include the accept/reject/partial annotation. After all revisions are complete, each agent states its revised confidence level and notes where its position now converges with or diverges from other agents' positions.
Produce the Integrator's Convergence Assessment: The Integrator maps the final state of agreement across three zones. Consensus Zone: Claims all three agents agree on, with the substance of the agreement stated explicitly. Partial Consensus Zone: Claims two of three agents agree on, identifying which agent dissents and the nature of the dissent. Disagreement Zone: Claims where all three agents differ, with each agent's position characterized and the underlying reason for non-convergence explained (different assumptions, different evidence weighting, different risk tolerance, genuinely undecidable given available information). For each dispute in the Disagreement Zone, assess whether it is resolvable with additional evidence or whether it reflects a legitimate values tradeoff that evidence cannot adjudicate.
Synthesize the completed debate trace into a final answer that follows from the consensus and partial-consensus claims, while explicitly noting which claims remain in dispute and what the user should weigh when deciding among competing positions. Show how the conclusion follows from the debate architecture rather than from a generic summary.
End the trace with a debate quality self-assessment: rate role differentiation (were the three agents genuinely distinct?), critique specificity (did critiques target specific claims or remain generic?), convergence honesty (were disagreements honestly characterized or forced into false consensus?), and overall robustness gain (did the debate produce a materially stronger answer than any single agent would have alone?).
Implementation Details
The Multi-Agent Debate framework, introduced by Du et al. (2023), establishes that multiple language model instances engaging in structured debate rounds can improve reasoning accuracy beyond what any single instance achieves alone. The core experimental finding is that debate produces convergence: across successive rounds, agent positions move toward consensus, and that consensus is more likely to be correct than the initial positions of individual agents. This effect is strongest when agents have heterogeneous reasoning stylesΓÇöa finding that directly contradicts the intuition that more identical agents voting produces better results. The mechanism is adversarial collaboration: each agent must defend its claims against specific, targeted critiques, creating a selection pressure that eliminates weak claims and strengthens surviving ones.
A critical finding from the MAD literature is the failure mode of homogeneous agent ensembles. When agents are identical or nearly identical in their reasoning patterns, their errors are correlated, and majority voting cannot correct those errors because the majority shares the same blind spot. Du et al. showed that current MAD frameworks sometimes fail to consistently outperform single-agent baselines precisely because homogeneous configurations produce mutual reinforcement rather than genuine adversarial testing. This finding is the rationale for this trace's mandatory heterogeneous role architecture: the Optimizer, Skeptic, and Integrator are not three copies of the same agent with different labels, but three agents with genuinely different cognitive priorities that force each to see what the others miss.
The A-HMAD (Adaptive Heterogeneous Multi-Agent Debate) architecture extends the basic MAD framework by introducing dynamic debate routingΓÇödecisions about which agents critique which others, the order of critique rounds, and when to terminate debateΓÇöbased on the characteristics of the specific problem and the quality of intermediate positions. Rather than a fixed, rigid debate structure, A-HMAD adapts the debate format to the problem: fact-heavy questions get more evidence-focused critique, creative questions get more possibility-expanding critique, and questions with clear normative dimensions get more values-surfacing critique. The Integrator's role in this trace embodies the A-HMAD principle by adjusting synthesis strategy to the nature of the disagreements it encountersΓÇöreconciling empirical disputes through evidence weighting, reconciling normative disputes through values surfacing, and characterizing genuinely undecidable disputes as such rather than pretending to resolve them.
Dynamic debate routing also governs the convergence assessment. In a traditional MAD setup, debate continues for a fixed number of rounds regardless of whether convergence has already been achieved, wasting computational effort. In this trace, the Integrator assesses after Round 2 whether positions have already converged on the key claims; if they have, Round 3 may be abbreviated to a confirmation round. Conversely, if Round 3 reveals that positions are diverging rather than convergingΓÇöa signal that the debate has surfaced a genuine values tradeoff or evidence gap that additional debate cannot resolveΓÇöthe Integrator characterizes the nature of the divergence rather than continuing to force additional rounds. This adaptive termination prevents both premature consensus (stopping before genuine disagreements are surfaced) and infinite debate (continuing when further rounds will not resolve the underlying dispute).
Output Information
Write the scratchpad as Markdown in root memory/{question_name}.md before responding to the user. Include this scale note near the top of the file: "Scale: default ΓÇö aim for around 150 numbered lines, or roughly 3,000 to 5,000 tokens, while adapting to the real complexity of the question and debate depth." Use numbered scratchpad items for scanability, but treat the number target as approximate and subordinate to usefulness. Keep the scratchpad public, inspectable, and concise enough per line that the debate structure remains easy to review. After writing the file, respond with the path, selected strategy and debate configuration, scale note, and final answer summary.
Specify Files And Length And Structure Of Output
Write the artifact to memory/{question_name}.md at the repository root, using the literal memory directory name. Structure the file with the sections Question, Strategy, Scale, Debate Configuration (listing the three agent roles and their epistemic postures), Round 1 ΓÇö Initial Positions (with subsections Optimizer Position, Skeptic Position, Integrator Position), Round 2 ΓÇö Structured Critiques (with subsections Optimizer Critiques, Skeptic Critiques, Integrator Critiques), Round 3 ΓÇö Revisions (with subsections Optimizer Revision, Skeptic Revision, Integrator Revision), Convergence Assessment (with subsections Consensus Zone, Partial Consensus Zone, Disagreement Zone), Debate Quality Self-Assessment, Synthesis, and Final Answer. The Scratchpad section should target around 150 numbered lines or roughly 3,000 to 5,000 tokens of public scratchpad detail, adjusted reasonably for very small or unusually broad questions. Use subsections inside the round sections when agent output becomes long enough that phases, branches, hypotheses, cases, or criteria would improve readability. If the user supplies a domain format, evidence source, or output constraint, preserve it inside this structure while keeping the default scale approximate.
Things Not to Do
- Do not produce identical or near-identical agent positions in Round 1ΓÇöif the Optimizer and Skeptic outputs are substantively interchangeable, the role differentiation has failed and the trace is invalid.
- Do not allow the Skeptic to produce generic critiques without specific claim targetsΓÇöevery critique must cross-reference a numbered claim from the target agent's position, and unspecific criticism must be discarded.
- Do not force false consensus by having agents agree in Round 3 when genuine disagreement persistsΓÇöunresolved disputes must be characterized honestly in the Disagreement Zone with an explanation of why convergence was not achieved.
- Do not skip roundsΓÇöeven if Round 1 positions appear to agree, Round 2 critique and Round 3 revision must be executed because surface agreement often hides latent disagreements that critique would expose.
- Do not present the Integrator as a tie-breaking vote rather than a synthesis agentΓÇöthe Integrator does not pick winners among competing positions but maps convergence, partial convergence, and divergence transparently.
- Do not allow the Optimizer to defer to the Skeptic's doubts without justificationΓÇöthe Optimizer may accept critiques that identify real weaknesses but must reject critiques that are unfounded or that would degrade solution quality without compensating benefit.
- Do not hide confidence changesΓÇöeach agent must state its confidence level in Round 1 and Round 3, and any significant shift must be explained.
- Do not omit the Debate Quality Self-AssessmentΓÇöthe trace must include an honest evaluation of role differentiation, critique specificity, convergence honesty, and overall robustness gain.
Input
The output of a multi-agent-debate-reasoning session: the selected strategy domain, the three agent role definitions, each agent's Round 1 position, all Round 2 critiques, each agent's Round 3 revision with accept/reject annotations, and the Integrator's convergence assessment. The trace generator consumes this structured output and produces the formatted audit trail. No additional user input is required beyond the raw session output.